To store for example two bits in a DRAM cell, it must be able to store four different voltage levels. A problem with such cells, is that noise margins are reduced to one-third that of a one bit per cell DRAM, which is too low to withstand the occasional .alpha.-particle hit.
A second problem with multi-bit storage cells relates to the method of sensing. No simple method of sensing has previously been designed, although attempts have been made to solve this problem, e.g. as described in the publication by M. Aoki et al, "A 16-Levels/Cell Dynamic Memory", ISSCC Dig. TECH. Papers 1985, pp 246-247, and in T. Furuyama et al, "An Experimental Two-Bit/Cell Storage DRAM for Macrocell or Memory-On-Application", IEEE Journal of Solid State Circuits, Vol. 24, No. 2, pp 388-393, April 1989. The technique described by Aoki cannot use normal sense amplifiers. It requires a precision analog D to A converter to implement a staircase waveform and a charge amplifier to sense data. The technique described by Furuyama requires the generation of precision reference levels to distinguish between four levels. These levels are not self-compensated for offsets developed in the sensing operation, and this method suffers from poor signal margin. Hidaka et al describe a technique for simultaneously reading two cells at a time in the article "A divided/Shared Bitline Sensing Scheme for 64Mb DRAM Core" in the 1990 Symposium on VLSI Circuitry 1990, IEEE, p. 15, 16 which while describing dividing a bitline, is not related to multiple bit storage in a single cell.
DRAMs have previously been built with cells holding up to sixteen bits of storage, e.g. in the aforenoted article by M. Aoki et al, for use in file memories. A 4 K test array is believed to have been the largest memory built using this design. Leakage characteristics of the DRAM cell were required to be very tightly controlled and even then, accurate sensing of the small voltage differences between levels becomes very difficult. Another problem with this scheme was the length of time required to access: a single read cycle required 16 clocks for the read followed by 16 clocks for the restore.
To implement a 2 bit DRAM, one can define the cell as storing one of four voltage levels V.sub.cell0, V.sub.cell1, V.sub.cell2 and V.sub.cell3, and reference voltage midpoints between these four voltage, which can be defined as V.sub.ref1, V.sub.ref2 and V.sub.ref3. These midpoints can be referred to, to differentiate between the four voltage levels. The relative voltage of these levels are shown in Table 1 below.
______________________________________ STORAGE REFERENCE ACTUAL VOLTAGES VOLTAGES VOLTAGE ______________________________________ V.sub.cell1 V.sub.DD V.sub.ref3 5/6 V.sub.DD V.sub.cell2 2/3 V.sub.DD V.sub.ref2 1/2 V.sub.DD V.sub.cell1 1/3 V.sub.DD V.sub.ref1 1/6 V.sub.DD V.sub.cell0 V.sub.SS ______________________________________
The storage voltages are the actual voltages stored in the cells, although the sensing voltages are somewhat more attenuated. Since sensing takes place on the bitlines which divide cell charge by the cell to bitline capacitance ratio, much lower voltages than those in the cell are actually sensed. In a standard DRAM, these voltage differences are in the order of 100-300 mV. It is the voltage midpoints between these smaller signals that must finally be generated to allow for correct sensing.
Furuyama et al in the article noted above describes one method of sensing these voltages. Furuyama et al used three sense amplifiers and three approximate midpoint sensing voltages. The cell charge is shared with the bitline, the bitline is split into three sections (sub-bitlines) and three sense amplifiers determine whether the cell charge is above or below their particular reference voltages. This data is then converted to two bits and a resulting output. Reconversion of the two bits allows approximate values to be driven into the bitline so that restore takes place after the read cycle. A write cycle operates in the same way as the restore section of the read cycle.
It should be noted that since the cell shares charge with three sub-bitlines, and the reference cell with only one sub-bitline, the reference voltage is about three times larger than it should be for sensing, casting doubt on the operability of this design. Secondly, three sense amplifiers are used, and since sense amplifiers have been growing proportionally larger and larger with each generation of memory, a minimum of sense amplifiers is desirable. A third problem is that the reference voltage is stored on a cell whose leakage does not track the leakage of the data cells, introducing another source of error into the circuit.